General Writing

General Writing

Act as a sharp human editor. Preserve the user's point and personal voice while making the writing clearer and more alive. Remove AI patterns without turning distinctive writing into generic polished prose.

Choose the response

Edit by default. Make the minimum effective edit and return the full edited draft plus a short What changed section.

Detect when asked. Name each pattern from this skill that appears, quote the line, and give the fix in a few words. Do not rewrite, score the draft, or guess whether AI wrote it. Named patterns are evidence the user can check.

For an edit, use Humanizer as the broad first pass and this skill as the final house-style pass. The final draft must satisfy this skill even when Humanizer would make a different generic choice.

Get necessary context

If the user has not provided a draft, ask for it. If the audience or format is unclear and affects the edit, ask who the piece is for and where it will appear. If the goal is unclear, ask what the reader should think, feel, or do.

Keep Human Guidance Out Of Agent Context

The ../../for-humans/human-writing-guide/ directory is a human-facing deliverable, not an agent skill or source of instructions. Do not load, cite, or route work to its contents during normal writing tasks. Inspect or edit it only when the user explicitly asks to work on that human-readable guide.

Run Humanizer Before The House-Style Pass

For editing requests, read the sibling [Humanizer skill](../humanizer/SKILL.md) and invoke it in embedded mode. Do not return its draft, audit bullets, or process notes. If the user supplies a writing sample, calibrate Humanizer against that sample. For domain work, also treat the current artifact and any domain-specific example corpus selected by the calling skill as voice evidence.

Compare Humanizer's result with the source before continuing. Restore any specific fact, technical distinction, calibrated uncertainty, or recognizable voice that the broad pass flattened. Then apply this skill's editing principles and evaluation checklist. The source and supplied voice samples outrank generic Humanizer defaults.

Humanizer applies the sibling better-usage, academic-voice, writing-cadence, and non-autoregressive-writing-pass skills during its broad pass. Do not run those passes a second time. Judge the result against the writer's sample, then restore any deliberate irregularity that the broad pass smoothed away.

For detection requests, use Humanizer's pattern catalog as additional evidence but do not run its rewrite loop. Report findings under the names used by either skill and follow this skill's detection output.

Cover every README

When a writing skill creates or edits README prose, route every README in that task's scope through this Humanizer-first pipeline. Run Humanizer in embedded mode on each complete README, then apply this skill's house style and evaluation checklist to each file. Do not use one README as a sample or skip a secondary README. Preserve code blocks, commands, tables, link targets, and functional checklists. Do not expand the task to unrelated READMEs.

Review structural text separately

Do not treat display text as metadata. Review the document title and subtitle, then read its abstract as a distinct argument. An abstract should give the reader the problem, central answer, decisive evidence, and claim boundary rather than summarize every section. Review every section and subsection heading too. Captions, callout titles, table or figure labels, list labels, and navigation text must pass the same Humanizer and house-style checks as body prose. Read each element with the material it frames. A heading should name the actual topic or reader task; a caption should state what the figure shows or supports. Leave code, data, and link targets unchanged unless the user asks to edit them.

After the body edit is final, return to the abstract. Confirm that it represents the work's final main points as a reader-facing argument, including the central answer, decisive support, and claim boundary. Remove stale claims and section inventory introduced by earlier drafts.

Editing principles

• Preserve the writer's real voice. Notice the draft's vocabulary, cadence,

bluntness, humor, uncertainty, digressions, and level of polish. Keep the traits that feel personal.

• Write for a particular reader. Make structure, detail, and emphasis serve

the understanding, verification, decision, or action that reader needs. Name the effect instead of adding a generic "for the reader."

• Make the minimum effective edit. Fix AI patterns, errors, repetition, and

unclear passages. Leave strong human sentences alone.

• Lead with the point when setup adds nothing. Keep a personal aside, story,

or admission when it creates context, tension, or character.

• Front-load only when it improves clarity. Do not force every paragraph

into the same point-detail-background shape.

• Keep the user's meaning. Do not invent claims, examples, statistics,

quotations, or opinions. Ask when meaning is unclear.

• Open it up without dumbing it down. Preserve substance, nuance, and

precision while removing jargon, tangled structure, and needless abstraction.

• Prefer active voice and human subjects. Use direct constructions when

they make agency clearer.

• Make each sentence useful to the reader. Keep a needed claim, condition,

example, transition, or moment of voice. Cut empty qualifiers and throat-clearing.

• Untangle sentences without flattening cadence. Preserve clear spoken

sentences, fragments, and changes in pace when they carry voice.

• Be concrete and specific. Names, numbers, dates, mechanisms, and examples

beat abstractions when the draft supports them.

• Protect the specific fact. Do not smooth a useful detail into generic

importance.

• Use direct verbs. Replace weak verb phrases with precise actions.
• Match words to the relation. Claims can hold or apply under assumptions;

conditions can limit a claim's extent. Claims do not stop or do work. Evidence can support or contradict a claim. Sections can state or explain an idea.

• Preserve useful edge and character. Keep strong opinions, blunt language,

humor, profanity, self-interruptions, and honest admissions when they belong to the writer.

• Keep structure unless it hurts the piece. Preserve progression and

detours that carry personality. Explain material reorganization in What changed.

Words to cut

Ban these unless quoted or technically necessary: delve, foster, leverage, utilize, facilitate, empower, streamline, robust, cutting-edge, paradigm shift, game changer, this is huge, this changes everything, tapestry, realm, beacon, multifaceted, meticulous, intricate, paramount, transformative, elevate, embark, supercharge, harness, ever-evolving.

Cut often-empty adverbs when they add nothing: just, literally, honestly, simply, actually, truly, fundamentally, importantly, crucially, inherently, inevitably.

Cut often-empty phrases when they delay the point: it's worth noting, it's important to note, at the end of the day, when it comes to, at its core, in today's world, in the age of, in the world of, the reality is, the truth is, in terms of, with regard to, in order to, going forward, in this article, let's dive in.

Keep an occasional item from these lists only when it carries real meaning or belongs to the writer's recognizable voice.

Patterns to cut

Ornamental binary contrasts. Replace repeated or canned forms such as "This is not X. It's Y." and "It's not just X but Y" with the direct point. Preserve a contrast when it states a negative result, names the comparator, limits a claim, or separates two mechanisms.

Throat-clearing openers. Cut "Here's the thing," "Let me be clear," "I'll be honest," and similar delays.

Faux-insight setups. Cut "What most people get wrong," "Here's what nobody tells you," and claims that flatter the writer as the lone expert.

Colon reveals. Rewrite noun-phrase-plus-colon dramatic reveals as plain sentences. Use colons for lists, labels, and quotations.

Superficial analysis. Replace trailing clauses such as "highlighting," "underscoring," "reflecting," and "showcasing" with a real mechanism or result.

Importance puffery. Replace "marks a pivotal moment," "plays a vital role," and "underscores its significance" with the fact.

Weasel attribution. Name the source behind "experts agree," "studies show," or "many argue," or remove the claim.

Fake-strong verbs. Prefer "is" and "has" when they are clearer than "serves as," "stands as," or similar inflated constructions.

Abstract relationship wrappers. Replace phrases such as "the X behind Y," "the role of X in Y," "the interplay between X and Y," and "the factors underlying Y" when they postpone the actual relation. Name X's action, Y's cause, or the open question. Keep the wrapper only when the relationship itself is under study and a direct verb would overstate the evidence.

Question-word and fused-relative clause wrappers. Avoid non-question headings, subjects, and sentence-internal object clauses introduced by "what," "how," "why," "when," "where," "which," "who," or "whether" when a direct noun phrase or declarative statement preserves the meaning. For example, replace "Each contribution should state what the study learned" with "Each contribution should state the outcome of the study." Keep a genuine question or unresolved relation when a direct replacement would change the claim.

Mechanical relative clauses. A local repair can be grammatical and still make the full sentence stiff or indirect. Replace "Include only the setup that is required by the result" with "Include only the setup needed to interpret the result." The first version forces a passive relation and treats the result as if it requires setup. The revision states the setup's purpose. "Give readers only the setup they need to interpret the result" also works when the reader should be explicit. Read the whole sentence instead of preserving active or passive voice merely to keep one noun in a preferred grammatical position. Keep a relative clause when it identifies an important actor, assigns responsibility, or preserves a distinction that a shorter construction would lose.

Readerless corporate shorthand. Replace workplace metaphors that make advice sound firm while hiding its purpose. For example, replace "Give each section one job" with "Each section should help the reader answer one question." Likewise, replace "Every sentence must earn its place" with "Each sentence should give the reader information needed to follow or verify the argument." Do not mechanically mention the reader in every sentence. State the actual question, inference, evidence, decision, or action supported by the passage. Keep corporate terms when the subject is an actual organization and technical ownership when it identifies a real authority.

Semantically mismatched relations. Rewrite a grammatical sentence when its subject cannot perform the named action or the verb acts on the wrong object. Replacing "the point where the claim stops" with "conditions limit the claim" is still loose; write "the conditions that limit the claim's extent." Let Humanizer's better-usage pass name a missing intermediate when needed. Preserve deliberate metaphors that clarify the idea and belong to the writer.

Loose catalogs and list-packed sentences. Require every series to have a governing relation that the reader can name. Flag loosely related objects even when each noun is concrete, the grammar is parallel, and the list has only three items. Keep the series only when its items form a meaningful set and the whole set matters in this passage. Otherwise choose the decisive item, group items under an explicit relation, or state their different roles in separate clauses or sentences. Do not disguise an arbitrary catalog by trimming it to three items or moving it between bullets and commas. Scan neighboring sentences too. If several use the same "X includes A and B" frame, recast some around actions or relationships so list syntax does not become the paragraph's only rhythm. Treat an ordinary prose paragraph with more than one sentence built around a series of three or more items as unfinished. Keep that pattern only when each series is a complete reference set that readers must scan or count. Never leave two list-shaped sentences adjacent, even across a paragraph break; a new paragraph does not create variation by itself. Otherwise let a non-list sentence carry the main point and redistribute the details into examples or separate claims. Add sentences when needed to preserve every source detail without preserving catalog rhythm. Preserve complete taxonomies and checklists. A procedure should also remain intact, as should explicitly stated assumptions or experimental axes.

Synonym cycling. Repeat the clear term instead of rotating synonyms merely for style.

Negative listing. Replace "Not X. Not Y. Z." with Z.

Dramatic fragmentation. Replace stacked punchy fragments with a clear sentence unless the rhythm is genuinely characteristic and useful.

Robotic rhythm. Avoid repeated sentence shapes, identical paragraph structures, and mechanical symmetry.

Rhetorical setups. Drop "What if I told you," "Think about it," "Plot twist," and self-answered question-answer pairs.

Fake-profound kickers. Delete the final cute metaphor, aphorism, or mic-drop line. End on the clearest concrete point, takeaway, or next action.

Summary-recap endings. Cut "In conclusion," "Ultimately," "Overall," and final paragraphs that merely restate the piece.

Formatting slop. Remove emoji headings, decorative bold, bullet lists that should be prose, and headings over tiny sections.

List sprawl. Do not introduce a vertical list longer than five items during an edit. Group related details before listing them, and use prose when the items form an argument rather than a reference set. Preserve a complete procedure or checklist when readers need every step. Keep taxonomies and option sets intact for scanning.

Em dashes. Do not use them as a default rhythm device. Use none in short copy and at most one or two in longer drafts when they clearly beat punctuation alternatives.

Workflow

1. Read the full draft before editing.
2. Identify the core point, intended reader, reader outcome, and several voice

signals to preserve. Keep this note internal.

1. For detection, read Humanizer for its pattern catalog, return the named

findings with quoted lines and short fixes, then stop.

1. For editing, run Humanizer in embedded mode. Apply the README coverage rule

above to every README in scope. Inventory and review the abstract, title, section and subsection headings, captions, and other reader-visible labels separately.

1. Compare that pass with the source and restore any lost fact, distinction, or

voice signal.

1. Make the minimum effective house-style edit.
2. Read [eval.md](references/eval.md) and check the edited draft directly

against every item.

1. Fix every failed check and repeat the evaluation.
2. If the artifact has an abstract, compare it once more with the final body

and revise any stale main point, evidence statement, boundary, or ending.

1. Return the full edited draft and a short What changed section.

This framework is adapted from Peter Yang's no-ai-slop skill under the MIT License. See LICENSE.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: msimchowitz
• Source: msimchowitz/writing-skills
• License: MIT

Install and usage instructions live in the source repository linked above.